| 摘要: | 網路聲量偵測是在市場調查時常使用的手法之一，常見手法是將某些事物被提及的次數作為熱門預測的指標。然而，只利用提及次數很難給予被提及的事物正確的評價；是否熱門往往牽扯到對於該事物的評價，因此本論文希望從社群網路的資料中找出目標事物的同時，判斷評論者對於目標的評價。
本論文主要使用多任務學習模型架構 (Multi-task Learning, MTL) 去進行模型上的設計，分別針對中文藝人的命名實體辨識(Singer Name Recognition, NER)和運用基於面向的情感分析(Aspect-Based Sentiment Analysis, ABSA)，針對目標做情感分析兩個任務去做研究。在NER任務中，利用多任務學習架構去取代常見的條件隨機域層 (conditional random field, CRF) ， 並在訓練資料中加入中文斷詞的資訊入中文分詞的相關資訊，讓模型在使用字向量的同時也能學習到詞方面的訊息，借此提升NER斷詞的準確度。在ABSA的任務中，我們在上一個NER預測模型的架構上，再加入情感判斷的任務，希望模型能在NER擷取同時做到目標的情感判斷。
本研究使用的資料為利用客製化爬蟲程式從社群網站上擷取之文章作為訓練資料，測試資料同樣從社群網站上隨機挑選文章，作為基準效能以評估模型之效能。實驗結果顯示，在NER階段，我們的藝人辨識模型在擷取未知藝人(OOV)效能達60%的Recall及50%的F1。在目標情續分析的任務中，我們延續 NER 階段所使用的模型架構，並在其架構上加入目標情緒分析任務做多任務學習，希望模型在找出目標實體的同時，給予目標實體情緒標記，而實驗結果顯示，在 NER 的結果中 F1 達到 88% 的，而目標情緒的部分有 56% 的 F1。;Network mentions is one of the methods commonly used in market research. The common method is to use the number of times certain things are mentioned as indicators of popular predictions. However, it is difficult to give the correct evaluation of popularity of the target entities only by the number of mentions; whether it is popular often involves the evaluation of the thing, so this paper hopes to find the target entities from the data of the social network and judge Commenter′s evaluation of the goal.
This paper mainly uses the multi-task learning model architecture (Multi-task Learning, MTL) to design the model, respectively for the name entity recognition of Chinese artists (Singer Name Recognition, NER) and the use of aspect-based sentiment analysis (Aspect-Based Sentiment Analysis, ABSA), to do two tasks of sentiment analysis to the target to do research. In the NER task, we use a multi-task learning architecture to replace the common conditional random field (CRF), and add chinese word segmentation information into the training data. So that the model can learn word information by using word segmentation, and improving performance of the precision of NER. In the task of ABSA, we add the task of sentiment analysis to the architecture of the previous NER prediction model. We hope that the model can judge the sentiment toward the target which is extracted by NER model.
The data used in this study retrieve from the community website using the customized crawler program as training data, and the test data is also randomly selected from the community website as the benchmark performance to evaluate the performance of the model. The experimental results show that in the NER stage, our artist identification model captures 60% of Recall and 50% of F1 of unknown artists (OOV). In the task of target sentiment analysis, we continue the model architecture used in the NER stage, and add target sentiment analysis tasks to its architecture to do multi-task learning, hoping that the model will give the target entity emotional markers while finding the target entity. The experimental results show that in the NER results, F1 reaches 88%, and the target emotion part is 56% F1. |
